SAM2-Labeling-tool

A simple labeling tool to help annotate frames in video data for segmentation tasks. It is based on SAM2 model: the user is prompted to sparsly annotate a set of randomly selected images, and this annotation is propagate through the dataset, using SAM2.

Overview

This repository provides a semi-automatic labeling pipeline for video and image datasets using the SAM2 segmentation model. It enables users to:

• Split videos into frames
• Annotate images interactively with foreground/background points
• Generate segmentation masks using SAM2
• Export datasets in LabelMe format for downstream tasks

Features

• Interactive annotation: Collect object/background points via mouse clicks
• Batch processing: Split videos and label images in bulk
• Automatic mask generation: Use SAM2 to create segmentation masks
• LabelMe export: Convert masks and annotations to LabelMe JSON format

Installation

1. Clone and install SAM2 dependencies

   git clone https://github.com/facebookresearch/sam2.git && cd sam2
   
   pip install -e .

2. Clone the repository:

   git clone https://github.com/Oassal/SAM2-Labeling-tool.git

3. Install dependencies (recommended: use Anaconda):

   conda create -n SAM2 python=3.10
   conda activate SAM2
   pip install -r requirements.txt

   • Required packages: numpy, matplotlib, opencv-python, Pillow, torch
   • You also need the SAM2 model weights and config files (see below)

Usage

1. Split Videos into Frames

Run the main pipeline with your video directory:

python main.py --video_dir "path/to/videos" --frames_per_split 100 --objects_names object1 object2 object3 ...

• --frames_per_split is the sampling value for each sequence, this value depends on the used GPU
• This will split each video into folders of frames.

2. Annotate Images

The script will prompt you to annotate selected images in each split folder. Use mouse clicks:

• Left click: Add object point
• Right click: Add background point
• Press 'n': Set current points as default for the object
• Press 'd': Clear all points
• Press 'enter': Save and close annotation

3. Generate Masks

After annotation, the pipeline will automatically generate segmentation masks using SAM2 for each split folder.

4. Export to LabelMe Format (Optional)

Add the --labelme flag to export the dataset in LabelMe format:

python main.py --video_dir "C:\path\to\videos" --frames_per_split 100 --objects_names object1 object2 object3 ... --labelme

File Structure

• main.py: Main pipeline script
• utils_labeling.py: Annotation and mask generation utilities
• labelme_utils.py / labelme_dataset.py: LabelMe export utilities
• utils_folders.py: Folder management utilities
• sam.py: SAM2 mask generation logic

Model Files

• Place your SAM2 model weights and config files in the project directory:
  • sam2_hiera_large.pt
  • sam2_hiera_l.yaml

Example Workflow

1. Place your videos in a folder
2. Run the main script with appropriate arguments
3. Annotate images as prompted
4. Masks and LabelMe datasets will be generated in each split folder

To be done

• Possibility of using other models
• Possibility of continuing labeling where you left off ...

License

MIT License